# OpenAI Unveils 722 AI-Generated Math Papers on GitHub Amid Skepticism
## A Landmark Release or Premature Hype?
OpenAI has dropped a massive collection of 722 mathematical manuscripts on GitHub, all generated by an internal AI model the company has not yet made available to the public. The documents are organized into 372 distinct families of related results, each family potentially bundling a central theorem alongside supporting arguments, alternative proofs, and related consequences.
The company states it submitted approximately 4,000 problems to the model and selected the outputs it considered significant enough for publication. On average, each result consumed the computational equivalent of roughly three hours of ChatGPT Pro thinking time.
## What’s in the Collection?
The breadth of the release is undeniable. The manuscripts span a wide range of mathematical territory, from incremental advances within established research programs to genuinely surprising breakthroughs. Most of the problems fall into categories that experienced mathematicians would classify as meaningful but not revolutionary.
Only one result has the potential to be truly transformative: a proof of the Quasi-Riemann Hypothesis. That single claim alone would represent a monumental achievement in the field if verified.
## The Formalization Gap
Here’s where serious concerns emerge. Of the 722 manuscripts, just 162 — roughly 22% — include a computer-checked main result formalized in Lean, a proof assistant that mechanically verifies every logical step. OpenAI has acknowledged that “some of the unformalized results could have issues,” leaving a substantial portion of the collection without reliable verification.
A passing Lean check only confirms that the proof follows logically from the statement as it was written in the formal system. It does not prove that the statement correctly captures the original problem, nor does it establish novelty or significance. The hard work of mathematical evaluation still lies ahead for human researchers.
## The Verification Debate
The reaction from the mathematical community has been sharply divided.
Andrew Sutherland, a mathematician at MIT, voiced strong skepticism. “Until and unless they release the model and people can replicate their results, I think you should treat any claims about one-shotting problems with a single agent as unverified,” he said, urging the community to “ask for receipts.”
The Institute for Advanced Study in Princeton weighed in with a statement expressing concern about the erosion of human understanding in mathematical scholarship. “We believe that human understanding of mathematics remains of paramount importance,” the statement read. The Institute had formed an advisory group that recommended OpenAI disclose the model name, the exact prompts, a summarized chain of thought, the time taken, and the compute cost for every result. OpenAI released average compute figures and reasoning summaries for 10 results but did not disclose the prompts.
Others, including Professor Abhishek Saha, expressed excitement, calling it “a very big day for mathematics,” while noting the limitations of what has been achieved so far.
## Transparency and Access Issues
The GitHub repository itself has drawn criticism. Researcher Keith Adler pointed out that the repository has issues turned off and has never accepted a pull request — an unusual stance for a project that explicitly asks for Lean formalizations from the community. “If you publish 722 manuscripts and ask for Lean formalisations, you need somewhere for people to send them,” Adler noted.
OpenAI has said it will add Lean formalizations as they become available, but the pace and scope remain unclear.
## A Different Path: Anthropic’s Approach
The release draws sharp contrast with Anthropic’s approach last month. The company published a Lean-checked proof of Fermat’s Last Theorem with all 13 million lines of code publicly accessible on GitHub. Unlike OpenAI’s claim of new results, Anthropic’s work formalized a theorem Andrew Wiles had already published in 1995 — a different but equally significant contribution to the field of formalized mathematics.
## What Makes This Release Different?
OpenAI distinguishes its effort from previous AI-assisted mathematics by claiming a largely single-agent workflow. An OpenAI spokesperson stated that nearly all manuscripts came from a single prompt handed to a single AI agent, though some may have required multiple attempts. The company also notes that the average compute cost per result was significantly lower than the Navier-Stokes claim from the previous month, which involved 10,000 coordinating agents working for 88 hours.
—
## Frequently Asked Questions
**What exactly did OpenAI publish on GitHub?**
OpenAI released 722 math manuscripts organized into 372 families of related results. The documents were generated by an unreleased internal AI model and cover a broad range of mathematical topics.
**Are these proofs verified?**
Only 162 of the 722 manuscripts — about 22% — include a Lean-formalized main result. Lean is a proof assistant that mechanically checks logical steps, but it does not guarantee that the result is new, important, or correctly addresses the original problem.
**Is this a breakthrough for AI in mathematics?**
It depends on who you ask. The scale of the release is unprecedented, but many mathematicians remain skeptical until the model is publicly available and results can be independently replicated. Only one result — the Quasi-Riemann Hypothesis — has the potential to be truly groundbreaking.
**Why are mathematicians skeptical?**
Skepticism stems from several factors: the model has not been released for independent testing, the prompts used were not disclosed, over 75% of the manuscripts lack formal verification, and some results have been described as incomprehensible even to trained mathematicians.
**How does this compare to Anthropic’s recent AI math proof?**
Anthropic published a Lean-checked formalization of Fermat’s Last Theorem with all 13 million lines of code publicly available. Unlike OpenAI’s release, Anthropic’s work did not claim new results but rather translated an existing proof into a machine-checkable format.
**Will OpenAI release the model eventually?**
OpenAI says it is still working to release the model responsibly, without providing a specific timeline. The company also plans to add Lean formalizations to the repository as they obtain them.
**How much compute did this take?**
On average, each result used the equivalent of roughly three hours of ChatGPT Pro thinking compute. This is notably less than other large-scale AI math efforts, such as the 88-hour, 10,000-agent Navier-Stokes attempt.
—
## Conclusion
OpenAI’s release of 722 AI-generated math manuscripts represents a bold experiment at the intersection of artificial intelligence and pure mathematics. The sheer scale of the output is remarkable, and the prospect of AI contributing meaningfully to mathematical research is undeniably exciting. However, the absence of full verification, withheld prompts, an unreleased model, and a divided response from the mathematical community all point to a long road ahead before this kind of work can be considered a legitimate addition to the scholarly record.
The tension between accelerating discovery through AI and maintaining the rigorous standards of mathematical proof has never been more visible. As the community digests these results and awaits the model’s release, one thing is clear: human mathematicians will be busy for a very long time.
Thank you for reading
